LocalScore test 4008: Apple M3 Ultra 20P+8E+60GPU
By anonymous community submission · LocalScore
- Source published
- Aug 9, 2026
- Test performed
- Aug 9, 2026
- Added to MiniPC.guru
- Sep 17, 2026
- Source reviewed
- Sep 17, 2026
- Tested hardware
- Apple M3 Ultra, twenty performance and eight efficiency cores, which is the 28-core processor · Apple M3 Ultra 60-core GPU, recorded by the tool as the accelerator · 96 GB unified memory · not applicable: token generation does not read from disk once the model is resident
- Software / version / OS
- llamafile · llamafile 0.9.2, commit a30b324 · Darwin kernel 25.6.0, arm64, reported by the tool as Cosmopolitan 3.9.7 MODE=aarch64 with board identifier T6031, which is the M3 Ultra. LocalScore records the chip rather than the machine, so for most Apple models that leaves the chassis unproven; here it does not, because the Ultra chip has only ever shipped in the Mac Studio
- Power profile
- not applicable: macOS exposes no selectable power profile on this machine
- Test conditions
- Llama 3.2 1B Instruct, 1.5 billion parameters, quantisation Q4_K Medium. This is a very small model and the figure is not representative of what a large one does on this machine; read it as an upper bound. The published value is the aggregate generation throughput across the tool's nine fixed prompt and generation lengths; the individual results are 191 tok/s at prompt 1024 with 16 generated, 126 at 4096/256, 161 at 2048/256, 156 at 2048/768, 176 at 1024/1024, 147 at 1280/3072, 192 at 384/1152, 208 at 64/1024 and 200 at 16/1536. Aggregate prompt processing 4394 tok/s, time to first token 261 ms
Editorial review checks the source and documentation; it does not certify the measurement or mean we tested the device ourselves.
